Non-white noise in fMRI: does modelling have an impact?

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Statistical Analysis

Background:

• Non-white noise in Blood Oxygenation Level Dependent (BOLD) functional magnetic resonance imaging (fMRI) arises from hardware imperfections, physiological processes (respiration, cardiac pulsation), and movement artifacts.
• This temporal autocorrelation in fMRI signal residuals invalidates standard statistical analyses due to violated independence assumptions.
• Current methods for noise reduction include high-pass filtering for low-frequency drift and autoregressive (AR) modeling for other noise sources.

Purpose of the Study:

• To introduce and evaluate Nuisance Variable Regression (NVR) as an alternative approach for mitigating non-white noise in fMRI data.
• To confirm the spatial distribution of fMRI noise sources using the General Linear Model (GLM).
• To assess the impact of NVR on temporal autocorrelation and normality of fMRI residuals and compare its performance against existing methods like SPM2's whitening approach.

Main Methods:

• Implementation of Nuisance Variable Regression (NVR) by incorporating confounding effects into a General Linear Model (GLM).
• Spatial mapping of fMRI noise sources to verify known distributions.
• Application of diagnostic statistics to evaluate the reduction in autocorrelation (first and higher order) and non-normality in residuals after NVR.
• Comparative analysis of NVR against the whitening method in SPM2.

Main Results:

• The spatial distribution of fMRI noise sources identified by NVR aligns with previously reported findings.
• NVR effectively reduces both first and higher-order temporal autocorrelation in fMRI residuals.
• The method also mitigates non-normality in the residuals, enhancing the validity of statistical inferences.
• NVR demonstrates comparable or improved performance relative to the SPM2 whitening approach.

Conclusions:

• Nuisance Variable Regression (NVR) offers a robust method for addressing non-white noise in fMRI data.
• By reducing autocorrelation and improving the distributional properties of residuals, NVR enhances the reliability of statistical analyses in BOLD fMRI.
• NVR presents a valuable alternative to conventional noise correction techniques, potentially improving the accuracy of neuroimaging research findings.